Recombinant Uncertainty in

Technological Search
Lee Fleming
Harvard University, Graduate School of Business, Morgan Hall T97,
Boston, Massachusetts 02163
lfleming@hbs.edu

hile the course of technological change is widely accepted to be highly uncertain

W and unpredictable, little work has identified or studied the ultimate sources and
causes of that uncertainty. This paper proposes that purely technological uncertainty derives
from inventors’ search processes with unfamiliar components and component combinations.
Experimentation with new components and new combinations leads to less useful inventions
on average, but it also implies an increase in the variability that can result in both failure
and breakthrough. Negative binomial count and dispersion models with patent citation data
demonstrate that new combinations are indeed more variable. In contrast to predictions,
however, the reuse of components has a nonmonotonic and eventually positive effect on
variability.
(Invention; Search; Recombination; Negative Binomial Dispersion Model)

Introduction 1991). While the assumptions of bounded rational-

While many have acknowledged the pervasive
uncertainty of technological change (Rosenberg 1996),
the ultimate sources of that uncertainty remain poorly
understood. Many scholars in the product life-cycle
tradition have observed that uncertainty peaks early
and decreases following convergence on a dominant
design (Anderson and Tushman 1990, Klepper 1997).
Radical and destabilizing change and the sources of
the life cycle are particularly difficult to predict and
often attributed to luck or individual genius (Tushman
and Anderson 1986, Ayres 1988, Mokyr 1990). Clark
(1985) proposes that technologies evolve and uncer-
tainty decreases as inventors iterate between the needs
of customers and the technological logic of their cur-
rent trajectory. Much research remains agnostic about
the causal sources of uncertainty and simply models
it as stochastic draws (Nelson and Winter 1982, Klep-
per 1996). Proponents of bounded rationality often
characterize technological change as an intrinsically
uncertain and, to varying degrees, blind search pro-
cess (Nelson and Winter 1982, Vincente 1990, March
ity may be more accurate than those of classical eco-
nomics, the metaphor of search remains informal and
was only recently developed empirically (Stuart and
Podolny 1996). Despite this widespread acknowledg-
ment of the importance of uncertainty, most research
gives the topic brief consideration en route to other
issues, and little work has attempted to identify and
empirically validate the causal sources of uncertainty.
Part of the difficulty in explaining the ultimate
source of uncertainty is the conceptual and empiri-
cal conflation of its many sources, and, in particular,
technological invention and commercial innovation.
There exist many causes of uncertainty in techno-
logical change besides purely technological sources,
including the adoption and diffusion of new technolo-
gies, market and customer acceptance, and competi-
tors’ strategic actions (Rosenberg 1996). Arguments for
intentional conflation of these sources notwithstand-
ing (Ruttan 1959), our confusion could be reduced
by returning to the classic differentiation between
technological invention and commercial innovation.
Schumpeter (1939) defined innovation as the com-

0025-1909/01/4701/0117$5.00 Management Science © 2001 INFORMS

1526-5501 electronic ISSN Vol. 47, No. 1, January 2001 pp. 117–132
 FLEMING
Recombinant Uncertainty in Technological Search
mercial application or adoption of an invention.
He argued that, “the making of the invention and
the carrying out of the corresponding innovation are,
economically and sociologically, two entirely differ-
ent things” (p. 85). While these sources remain inter-
dependent, their conflation confuses our explanations
of their respective processes and contributions to
uncertainty. To avoid such confusion, this paper will
focus on invention and purely technological sources
of uncertainty.
Attempts to understand the source of technological
uncertainty have also been frustrated by a lack of
understanding of the process by which inventors
create new technologies. A theory of the process
of invention would facilitate our understanding of
the sources of technological uncertainty. To address
this issue, this paper synthesizes two classic perspec-
tives on the sources of technological novelty: first,
invention is a process of recombination, and, second,
invention is indeed an inherently uncertain and, there-
fore, typically local search process.
Invention As a Process of
Recombinant Search
Many scholars have proposed that recombination pro-
vides the ultimate source of novelty (Gilfillan 1935,
Usher 1954). Schumpeter (1939, p. 88) observed that
“innovation combines components in a new way,
or that it consists in carrying out New Combina-
tions.” Nelson and Winter (1982, p. 130) state that
“the creation of any sort of novelty in art, science,
or practical life—consists to a substantial extent of
a recombination of conceptual and physical materi-
als that were previously in existence.” Henderson and
Clark (1990) argue that the mere rearrangement of
previously used components can itself cause destabi-
lizing industrial change. Hargadon and Sutton (1997)
describe how a design consulting firm creates nov-
elty by brokering ideas and technologies between their
clients’ industries.1 Basalla (1988) develops an analogy
1
While the concept of recombinant search as the source of novelty
may well apply outside the strictly focused realm of invention—
indeed, to cuisine, innovation (Abernathy and Utterback 1978,
Kogut and Zander 1992, Levinthal 1998), and process improvement
(Romer 1993)—I can only comment empirically on invention.
118
from Kroeber’s (1948) image of a tree of cultural
artifacts. Unlike genetic trees, however, the branches
of the tree of technology can fuse together. “Separate
types or branches fuse together to produce new types,
which merge once again with still other branches”
(Kroeber 1948, p. 138). Similar to the theory of natural
evolution before the discovery of DNA, the analogy
breaks down at the genetic level. “We who postulate
theories of techological evolution likewise have our
Darwins but not our Mendels” (Kroeber 1948, p. 210).
Although the idea of recombination as the source of
novelty has been widely discussed, the implications of
the idea remain undeveloped. Previous work suggests,
however (Schumpeter 1939, Henderson and Clark
1990), that an invention can be defined as either a
new combination of components or a new relationship
between previously combined components. While
knowledge, science, algorithms, culture, applications,
and manufacturing processes also influence invention,
they are not part of an actual artifact and are not
actually instantiated in an invention. They strongly
influence the process of inventive search, however,
by inspiring, aiding, explaining, or constraining the
use of particular components or combinations. Collo-
quial usage of the word “components” implies known,
available, and commercially available hardware. In
this paper, however, “components” will denote the
constituents of invention, along the lines of what
Schumpeter calls “factors” (1939, p. 88).
The Scope of Potential Recombination
Even though the number of potential components
overwhelms the imagination, there are no restrictions
on the scope of their recombination. Components are
not like genes and “similar” technologies are not
like species (Basalla 1988). In contrast to variation
processes within genetically isolated populations,
inventors can recombine any components within their
purview. Perceptions that certain technologies or com-
ponents “belong together” develop through social
construction and previous association. For example,
if an electrical engineer of the 1940s had been asked
about his profession’s use of sand and aluminum,
he probably would have replied with a blank stare.
Today, he or she probably would reply that they are
the most common basic materials of semiconductors
Management Science/Vol. 47, No. 1, January 2001
 FLEMING
Recombinant Uncertainty in Technological Search
and the focus of much research investment. Clearly, no
technology evolves independently of the entire world
of made things. At any point in technological evolu-
tion, any component is at risk of being recombined
with any other component. The made world evolves
as a holistic, continuous, and interdependent web, and
not as a disjoint assortment of separate trajectories or
product life cycles.
Investors constantly import previously untried
components from outside the extant made world,
for example, the use of medicinal substances from
tropical jungles or the exploitation of petroleum and
natural gas, substances that had little application to
technology prior to the 19th century. Investors also
create new components through encapsulation and
hierarchical modularization of component sets (Simon
1996, Baldwin and Clark 1999). Such “black box” engi-
neering efforts make the underlying components less
salient but do not destroy the recombinant potential
of the components or the encapsulating module. As
these processes make additional components available
for further recombination, the diversity of the “made”
world increases. Because all previously used compo-
nents and inventions provide potential constituents
for future inventions, the potential risk set for recom-
bination is the entire extant made world. Because
inventors also scan outside the made world, anything
not yet derived from the natural world can also be
considered as a potential constituent of invention, as
part of the theoretical risk set. Recombination usually
occurs, however, between components that are salient,
proximal, and available for the inventor.
These ideas prompt us to look backward in time
to consider the components and untried combinations
that were available at the time of invention; they
prompt us to look forward to predict which inven-
tions are more likely to motivate further recombi-
nation. Prediction of further recombination depends
on the fundamental tension between exploration of
untried possibilities and exploitation of previous suc-
cesses (March 1991).
Cognitive, Social, and Technological
Influences on Recombination
Because the agents of recombination are people,
the process of invention remains strongly influenced
Management Science/Vol. 47, No. 1, January 2001
by cognitive and social phenomena. The most
fundamental influence is a limitation on the num-
ber of potential components and combinations that an
inventor can simultaneously consider. Because every
invention can be incorporated in further recombina-
tions, inventors’ combinatoric potential has grown
explosively (Weitzman 1996). The set of potential com-
binations and, a fortiori, the possible ways that each
set of potential combinations can be combined has
become essentially infinite. It has become impossible
for individual inventors, groups, even entire commu-
nities of inventors to have more than an infinitesimal
understanding of all these potential combinations
and relationships. As a result of this combinatoric
explosion, inventors and their organizations and
communities must focus and recombine locally from
a limited set of components and combinations.
These arguments follow the assumption of bounded
rationality and local search (March and Simon 1958,
Nelson and Winter 1982, Cohen and Levinthal 1990,
March 1991, Kauffman 1993, Stuart and Podolny
1996). Localness corresponds to inventors’ familiar-
ity with their recombinant search space. Local search
or exploitation (March 1991) occurs when an inven-
tor recombines from a familiar set of technology
components or refines a previously used combina-
tion. A pastry chef searches extremely locally when
he mixes previously used dyes in a new propor-
tion to create a novel frosting color. Distant search
or exploration (March 1991) occurs in the opposite sit-
uation, when inventors try completely new compo-
nents or combinations. The early auto industry pro-
vides many examples of successful and unsuccessful
distant search including various power sources, pneu-
matic tires, brakes for each passenger, four-wheeled
diamond configurations, and, in the 1930s, a combined
car and plane configuration (Basalla 1988).
One would expect that local recombination is more
certain and, on average, more successful. To the
extent that inventors draw from familiar compo-
nent sets and refine previous combinations, they are
less likely to develop a completely useless invention.
They also decrease their upside potential, however,
of developing a radically different invention that is
of much greater impact. Local recombination is more
certain because inventors learn from past failures.
119
 FLEMING
Recombinant Uncertainty in Technological Search
Inventors learn which components failed in previ-
ous inventions and stop using them. They learn to
avoid the combinations and architectures that failed
in the past. They winnow and bound the less suc-
cessful regions of recombinant space (Vincenti 1990).
This winnowing and bounding improves the average
usefulness of inventive efforts but also decreases the
possibility of wildly successful inventions.
As inventors reuse components they begin to
understand and characterize those components. They
begin to understand which components are more
or less useful in different contexts. Such knowledge
enables selection and exploitation of more appropriate
components in future inventions. Understanding and
knowledge increase with use; the greater the use of a
component, the greater the knowledge of and famil-
iarity with it. Knowledge can also be forgotten, how-
ever, and is therefore more potent the more recently it
has been gained (Argote et al. 1990). More recent and
frequent usage therefore implies greater knowledge
and familiarity. Inventions that incorporate familiar
components should be more useful because inventors
can select more appropriate components. Because they
can better predict the performance of the included
components, incorporation of familiar components
should also decrease inventive uncertainty and, hence,
the variability of outcomes.
Inventors can draw on others’ knowledge and
experience in addition to their own. While social prox-
imity certainly increases the ease and likelihood of
sharing, knowledge about component use will still dif-
fuse between organizations (Allen 1977) and techno-
logical communities (Bijker 1987). Many mechanisms
facilitate knowledge diffusion, including person-
nel movement, personal friendships, organizational
merger, education, reverse engineering, technical liter-
ature, and strategic alliances (Ahuja 2000). Although
much knowledge will be lost or changed in its dif-
fusion, sharing will take place at all social levels,
between individuals, organizations, and communities.
Taken together, these arguments imply the following.
Hypothesis 1. Recombination of familiar components
will increase an invention’s usefulness.
Hypothesis 2. Recombination of familiar components
will decrease inventive uncertainty.
120
Parallel arguments hold for the refinement and
improvement of combinations. Engineers gain expe-
rience with particular combinations by using them.
As inventors learn which combination relationships
or architectures are less useful, they avoid those
approaches. Such learning and knowledge helps
them improve their inventive efforts on average and
decrease inventive uncertainty. As with knowledge
about individual components, knowledge about com-
bination relationships increases with frequency of use
and decreases since last use. As with individual com-
ponents, knowledge of combinations and their opti-
mal relationships will diffuse throughout the made
world. These arguments imply combination hypothe-
ses that parallel the individual component hypotheses.
Hypothesis 3. Refinement of familiar combinations
will increase an invention’s usefulness.
Hypothesis 4. Refinement of familiar combinations
will decrease inventive uncertainty.
Practitioners have long recognized the value of
reuse and refinement (Mead and Carver 1980) and the
difficulty of exploring new regions of the essentially
infinite design space. For example, Altschuler (1998)
explicitly recommends searching previous inventions
for universal analogies and possible applications to
new contexts, and Goldenberg et al. (1999) propose
and empirically validate a method to identify and
link profitable dependencies between previously used
components. Unfortunately for inventors, however,
these benefits of familiarity do not last forever. This
results from the technological and social-psychological
exhaustion of potential refinements, given a particular
combination. Both influences run counter to the posi-
tive effects of local search.
Technological exhaustion occurs because most of
the possible relationships between a set of compo-
nents have already been tried. As Kim and Kogut
(1996) argue, “The repeated application of a particular
set of technologies or organizing principles eventually
exhausts the set of potential combinations.” The argu-
ment generalizes Sahal’s (1985) demonstration of the
decreasing returns to physical scaling. For example, if
semiconductor inventors restricted their usage to their
original materials of aluminum and bipolar transis-
tors, progress in the field would have halted long ago.
Management Science/Vol. 47, No. 1, January 2001
 FLEMING
Recombinant Uncertainty in Technological Search
However, because inventors began using new materi-
als such as copper interconnect and new combinations
such as metal oxide semiconductors, semiconductor
chips have continued to shrink and the trajectory has
repeatedly avoided exhaustion.
Exhaustion in technological potential also stems
from social and psychological sources and can ulti-
mately be traced to inventors’ frames and “imagi-
nary life cycles” (Henderson 1995). If a community of
inventors accepts or reifies a particular combination as
mature, then they are less likely try new combinations
or broaden their component set. Henderson describes
three such attitudes in her study of the premature
obituaries for optical lithography in 1977: Inventors
failed to consider the possibility of new architec-
tures, improvement in their components, or changes
in users’ application of the technology. These tech-
nological and social-psychological effects run counter
to the improvement that results from familiarity and
learning.
Hypothesis 5. Cumulative use of a combination will
decrease an invention’s usefulness.
Methods
Thinking of invention as a recombinant search pro-
cess implies that an invention’s usefulness and uncer-
tainty can be predicted from previous usage of its
components and particular combination of compo-
nents. I test these ideas with patent data and estima-
tion of a statistical model. The models predict future
prior art citations to a given patent based on previous
usage of its assigned subclasses and particular combi-
nation of subclasses. The hypotheses place additional
constraints on the empirical work, however. If engi-
neers can recombine any technology with any other,
the data should include recombination across techno-
logical communities. If invention is a process of search,
learning, and exhaustion, the data should include the
history of component and combination use. Finally,
if the uncertainty of invention decreases with learn-
ing and the use of familiar components and combina-
tions, the statistical model should estimate the impact
of such learning and familiarity on the variability of
inventive outcomes.
Management Science/Vol. 47, No. 1, January 2001
Data
The unit of analysis is an U.S. patent granted dur-
ing May or June of 1990
n = 17264.2 The data came
from the MicroPatent (1996) product. While patent
data are admittedly imperfect records of technology
(Levin et al. 1987) and conditional on a successful
patent application, these data still represent a large
portion of failed and successful invention. Indepen-
dent measures (recombinant history) came from data
before July of 1990 and dependent measures (citations)
came from data after July 1, 1990. Citations are used
as controls and a dependent variable only—they are
not used to trace the path of combination reuse, learn-
ing, or the diffusion of knowledge. The design is cross-
sectional and does not consider multiple time periods.
While the sample is large enough to provide very sig-
nificant results for all the hypothesis tests, future work
should consider other time periods to minimize the
chances of idiosyncratic sampling.
I observe recombination across all technological
communities by looking simultaneously at all patents
granted during the two-month time period, thus sat-
isfying the first modeling constraint. If the data did
not consider all communities simultaneously, I would
need to define the boundaries between particular
technological communities. Because such boundaries
are transient, permeable, overlapping, and nested, it
would be empirically intractable to determine them
for even a fraction of the made world. Finally, two
hundred years of patent data satisfy the need to
observe recombinant history of component and com-
bination use.
Measures
Table 1 lists and describes all variables. For the depen-
dent variables of mean and variance, I measure an
invention’s usefulness as the number of prior art cita-
tions that it receives from subsequent patents. To be
granted a patent, the applicant must establish the nov-
elty of her invention relative to all previous inventions.
She establishes her claim to novelty by identifying
(almost always after the fact of invention) and citing
similar “prior art.” The patent examiner then reviews
and usually supplements these citations (Carr 1995).
2
I chose the months of May and June at random.
121
 FLEMING
Recombinant Uncertainty in Technological Search

Table 1 Description of Variables

Variable Type Description Measure

Citations
Mean technology
Variance technology
Number of prior art citations
Single-class dummy
Number of subclasses
Newest subclass
Number of classes
Component familiarity
Combination familiarity
Cumulative combination usage
dependent invention’s usefulness or importance prior art citations by future patents to focal patent
control expected citations to technically similar patents weighted fixed effects by focal patent’s sub-class
membership in class
control expected variance in citations to technically similar weighted fixed effects by focal patent’s sub-class
patents membership in class
control patents that cite more heavily should be more number of prior art citations by focal patent
heavily cited
control process of recombination is finer grained than can equals 1 if focal patent assigned to only one subclass
measure
control number of invention’s components number of focal patent’s subclasses
control artifact of patent classification system equals the minimum number of previous uses
amongst the focal patent’s sub classes
control breadth of patent classifications number of focal patent’s classes
independent inventor’s familiarity with components of the recent and frequent usage of focal patent’s subclasses
invention across all U.S. patents
independent inventor’s familiarity with particular combination recent and frequent usage of particular subclass com-
of components bination across all U.S. patents
independent cumulative number of inventive trials with exactly cumulative number of U.S. patents with exactly same
same combination subclass combination

Bibliometric studies have repeatedly demonstrated
that future prior citations to a patent correlate with
its technological importance and value (Albert et al.
1991, Hall et al. 2000). To make maximal use of the
data, I measure prior art citations to a focal patent for
6 years and 5 months after its granting. This period
should capture the bulk of citations to a patent as these
citations typically reach a plateau after about three
years from the grant date (Jaffe et al. 1993). While there
appears to be strong correlation between the rates of
early and later citations to a patent, researchers are
actively pursuing the topic (Hall et al. 2000).
To measure the independent variables, I proxy com-
ponents with patent subclasses. The patent office cat-
egorizes all patentable technologies into some 400
“class references”. Each class is also subdivided into
very fine divisions of technology or approximately
100,000 “subclasses” (Trajtenberg et al. 1997). The
patent office typically assigns each patent into mul-
tiple subclasses within and across major classes. The
patent office also establishes and updates new classes
and subclasses each year, as technology changes (Carr
1995, p. 128). This retrospective updating enables
historical consistency in the measurement of compo-
nents across time.
I do not propose that inventors recombine patent
subclasses directly, only that subclasses can be used
to observe indirectly the process of recombinant
search and learning. I will illustrate the correspon-
dence between components and patent subclasses
with patent 5,136,185, coauthored with John Walther.
It was (as of Dec. 1996) classified in four subclasses:
326/16 (with test facilitating feature), 326/31 (signal
sensitivity or transmission integrity), 326/56 (tristate
(i.e., high impedance as third state)), and 326/82
(current driving (e.g., fan in/out, off chip driving,
etc.)). Each of these subclasses corresponded to a
well-understood digital hardware component at that
time. Each of them can be found in (and, for this
patent, were drawn specifically from) contemporary
textbooks of digital design such as McCluskey (1986).
A test facilitate feature (p. 426) has become a neces-
sity with large computer chips due to their millions
of gates. To test the innards of a chip, the chip must
be operated in a “test mode.” A transmission gate
(p. 118) simply passes on a signal when enabled.

122 Management Science/Vol. 47, No. 1, January 2001


Recombinant Uncertainty in Technological Search
A tristate driver (back cover diagram of digital com-
ponents and pp. 105, 119, 142) is a gate that can drive a
bus or be turned off to present high impedance to the
bus. Finally, “fan in” and “fan out” (p. 104) refers to
the number of components that can drive or be driven
by a particular gate (high fan out basically means a
big gate). While we were not aware of the subclass
definitions at the time of the invention, we were very
familiar with these basic building blocks of our tech-
nological community, namely digital hardware engi-
neers. I propose that subclasses serve as proxies for
these building blocks.
With the subclassifications and date of application
of each patent (interpolated prior to 1975), I develop
the three independent variables for each focal patent:
component familiarity, combination familiarity, and
cumulative combination usage. Component familiarity
proxies inventors’ familiarity with components, based
on the average degree to which the components have
been recently and frequently used. The assumption
is that inventors will be more familiar with compo-
nents that have been recently and frequently used.
I first calculate an individual measure for each sepa-
rate subclass of the focal patent. To the extent that a
particular subclass has been recently and frequently
used, its individual measure will be higher. For each
focal patent, I look backward to 1790 and consider
each of its individual subclasses in turn. Whenever
a particular subclass has been used in any previous
invention, I multiply the indicator of occurrence by
an exponentially decaying component (1). This expo-
nential component represents the loss and forgetting
of knowledge. For example, it is more likely that an
inventor will have learned from previous use of a
subclass, if that sub-class was used three years prior,
instead of thirty. These occurrences are then summed
and averaged (2).
Individual component familiarity of patent i’s
subclass j ≡

count by the exponential component, to reflect the
Iij = 1 patent k uses subclass j
all patents k granted
before patent i
 application date of patent i− application date of patent k 
−
×e time constant of knowledge loss
Management Science/Vol. 47, No. 1, January 2001
FLEMING
Average component familiarity of patent

all subclasses Iij
 j of patent i
i s subclasses ≡ Fi =   (2)
all subclasses 1
j of patent i
I set the time constant of knowledge loss at five
years. This constant in the denominator of the expo-
nential implies that approximately one-third of the
knowledge remains after five years, or a yearly loss
rate of 18%. Argote, Epple, and Darr (Argote et al.
1990, Epple et al. 1991, Darr et al. 1995) have estimated
a much higher geometric loss parameter for manu-
facturing and service organizations, between 40% to
97% per year. It is unlikely that design technology
would experience such a high rate of loss, however.
Design knowledge is far less contextual and more eas-
ily articulated than manufacturing or service experi-
ence. Design knowledge is more likely to have been
recorded in trade journals, firm documentation and,
of course, patents. It may have been actually real-
ized in prototypes and products and is far more
likely to have required substantial personal effort and
investment on the part of its designers. All of these
influences would argue for a slower loss rate than
manufacturing and service organizations, but the pre-
cise estimation certainly constitutes a valid research
question.
I calculate combination familiarity and cumulative
combination usage similarly. The cumulative usage
measures how many times since 1790 a particular
combination of subclasses has been used. Combina-
tion familiarity proxies inventors’ familiarity with the
combination, based on the degree to which the com-
bination has been recently and frequently used. For
each focal patent, I consider its particular combina-
tion of subclasses. I then look back in time at all
other patents that used an identical combination of
subclasses. For each previous patent that used an
identical combination of subclasses, I increment the
cumulative combination usage (3). For the combi-
nation familiarity variable, I multiply each indicator
(1) loss and forgetting of experiential knowledge regard-
ing the combination (4). All the measures remain indi-
rect because they assume that learning occurs with use
 and that knowledge diffuses throughout technological
123

Recombinant Uncertainty in Technological Search
communities and the made world. Given that all the
independent variables are highly skewed, I took the
square root of their original value. This minimized
the effects of outliers and enabled more parsimonious
modeling.
Cumulative combination usage of patent i ≡ Ci =
 all patents k granted
1patent k uses identical
all patents k granted
before patent i
combination of subclasses
as patent i
Combination familiarity for patent i ≡ Ri =

1patent k uses identical combination
all patents k granted
before patent i
of subclasses as patent i 
 application date of patent i − application date of patent k

×e time constant of knowledge loss
Given the theoretical interest in recombinant search,
the models should ideally control for other differences
between inventions, for example, the differences in
citation patterns across different technology classes.
I calculate technology mean and variance by measuring
citations to a technology class prior to the dependent
measures period. I first calculate the expected num-
ber of citations to each class in the five and a half
years prior to July of 1990 (5). Equation (6) illustrates
similar calculations for the variance of each patent
class. Using these numbers for each class, I control
for the expected mean and variance of citations to
each focal patent based on a weighted average of its
assigned technology class references. I calculate this
by multiplying the proportion of a patent’s subclass
assignments within a particular class by the average
cites/class of that class. As an additional control of
differences in citation activity, I include the number of
prior art citations that a focal patent makes to previous
patents. Although none of these measures differences
across industry directly, in aggregate they provide
some control of nonrecombinant heterogeneity, to the
124
FLEMING
extent that industries have similar technologies and
citation patterns.
Average citations to a patent classified within class
i ≡ i =


1patent j cites a patent within class i
all patents i granted
from 1985−19905
∗
proportion of cited patent within class i


1patent k classified within class i
∗
proportion of patent k within class i
(3) from 1985−1990
(5)
Variance in citations to patents classified within class
i ≡ i2 =


1patent k classified within class i ∗
proportion
all patents k granted
from 1985−19905
of patent k within class i ∗
citesk − i 2


1patent k classified within class i
all patents k granted
from 1985−19905 ∗
proportion of patent k within class i
(6)
(4) Given the theoretical assumptions of bounded ratio-
nality and component search, the models should con-
trol for the number of the invention’s components.
These can be proxied by the focal patent’s number of
subclasses. Because it becomes increasingly difficult for
an exact match to occur as the number of compo-
nents increases, I include a second-order term for this
as well.3 Because the recombinant history of patents
with only one subclass remains unobservable, I also
include a dummy variable for single subclass patents
(they comprise 8.1% of the data).
Finally, the models should control for artifacts of
the patenting system. The patent office periodically
updates technology classifications and creates new
subclasses (these analyses use the classifications in
effect as of November 1996). Because such reclassifica-
tion essentially recognizes newly successful technolo-
gies, the first few patents of a retrospectively identified
technology stream are likely to be highly seminal.
Hence, patents that include a newly designated sub-
class will probably be highly cited. I include a newest
3
I would like to thank an anonymous reviewer for this suggestion.
In addition to adding the squared term, I checked the sensitivity
of the results by estimating the model separately on subsets of the
data, for patents classified in two and three subclasses. The results
agreed with those presented below.
Management Science/Vol. 47, No. 1, January 2001
 FLEMING
Recombinant Uncertainty in Technological Search

Table 2 Descriptive Statistics for All U.S. Patents, May/June 1990 ties of maximum likelihood estimators (Greene 1993,

n = 17264
p. 133).
Variable mean stnd.dev. minimum maximum

citations 380 488 000 8200 Negative Binomial Count Models

mean tech control 119 041 033 303 The dependent variable of citation counts takes on
variance tech control 465 265 036 2055 only whole number values (that is, 0, 1, 2, etc.). The
number of prior art citations 763 699 000 11000 use of a linear regression model on such data can
single subclass dummy 008 027 000 100 yield inefficient, inconsistent, and biased coefficient
number of subclasses 421 331 100 13000 estimates (Long 1997). Explicit count models can avoid
newest subclass 10707 10397 000 187400 these problems. Researchers often use Poisson models
number of classes 178 095 100 1200 to analyze count data, but Poisson models assume that
component familiarity 060 027 000 228 the mean and variance of the observed distribution are
combination familiarity 028 075 000 1125 equal. These data, like most count data, exhibit over-
cumulative comb usage 062 160 000 2676 dispersion—the variance is greater than the mean.
Negative binomial regressions explicitly accommo-
date this over-dispersion, however, by enabling the
subclass control variable that equals the minimum of
variance to be greater than the mean. Recent develop-
the least used subclass (for example, if a patent had
ments (King 1989, Jorgensen 1997) also support inde-
three subclasses and each subclass had been previ-
pendent estimation of effects on the mean and the
ously used in 56, 2, and 43 patents, the variable would
variance of the predicted mean, as functions of poten-
equal 2). This variable is very similar to the dependent
tially differing sets of variables.
variable and hence probably decreases the effects of
the independent variables. I also include the number of To begin with, consider a model that estimates the
classes to which a patent belongs to reduce the effect of mean number of citations that a patent should receive,
additional citations to a patent, similar to that which given its independent variables. The observed cita-
occurs when a scientific paper straddles fields. tions to the patent will not correspond exactly to
Tables 1, 2, and 3 list the variables, descriptive statis- the prediction, however, and will be distributed with
tics, and a correlation matrix. Combination familiarity some variance around the expected mean. I opera-
and cumulative usage demonstrate high correlation in tionalize the uncertainty of invention by estimating
Table 3. This is not a problem except for inflated stan- the effects of substantive variables on this variance.
dard errors, given the desirable large sample proper- For example, to the degree that some independent

Table 3 Correlation Matrix for All U.S. Patents, May/June 1990
n = 17 264

cites mean variance prior single numsub newest class comp combin

mean tech control 031

variance tech control 030 094
number prior art cites 012 002 000
single subclass −006 000 000 −007
number subclasses 011 001 007 008 −029
newest subclass −008 −008 −007 −003 024 −023
number of classes 008 −002 001 006 −024 051 007
component familiarity 016 030 035 −001 −009 027 001 014
combination familiarity 001 006 005 −004 047 −027 −005 −023 017
cumulative comb usage −005 −003 −003 −005 065 −030 −005 −026 006 0.85

Management Science/Vol. 47, No. 1, January 2001 125

 FLEMING
Recombinant Uncertainty in Technological Search

variable decreases inventive uncertainty, its effect 1/i . Integrating (10) with the density function of (11)
on the variance of this error distribution should be gives the probability of the negative binomial of (12).
negative. 
Most derivations of the negative binomial start from i i i−1
−
i i 
g

i  =
e for i > 0
a basic Poisson model (7). The basic Poisson model

i  i

estimates the probability of an observed count, condi- and

 = t −1 e−t dt (11)
0
tional on an expected mean i . To avoid negative (i.e.,  i  yi
undefined) expected values for the mean i , Poisson

yi + i  i i
Pr
yi xi  =  (12)
models typically parameterize explanatory variables yi !

i  i + i i +  i
as an exponential function (8). The method of maxi-
While the first term of (9) fully specifies the mean
mum likelihood is then applied to the joint frequency
of the negative binomial, various parameterizations
formed from the product of the marginal frequencies
of i remain possible. Cameron and Trivedi (1986)
of (7), to determine the coefficient values that are most
propose the negative binomial II parameterization (or
likely to result in the observed counts.
Negbin II model) when the variance/mean ratio of
e−i i i
y the observed data is linear in the mean. By contrast,
Pr
yi xi  =  (7) the Negbin I holds the variance/mean ratio constant.
yi !
The Negbin II specification is also much more robust
E
yi xi  = i = e
xi   (8) to distributional misspecification than other param-
eterizations (Cameron and Trivedi 1986). I verified
The negative binomial model replaces the Poisson
the applicability of the Negbin II by regressing pre-
mean i with the random variable ˜ i (9). This replace-
dicted counts on the quantity (residualsˆ2/predicted).
ment enables the inclusion of an error term
i = ei —
The coefficient and intercept were positive thus sup-
and allows the predicted mean to vary according to
porting the Negbin II parameterization (Cameron and
the distribution of the error term. Substitution of ˜ i
Trivedi 1986). Also consistent with Cameron and
for i in (7) results in (10).
Trivedi’s argument (1986), Negbin II models demon-
˜ i = e
xi +i  = i
i  (9) strated much more significant log likelihoods than
Negbin I models.
e−i
i
i
i yi
Pr
yi xi 
i  =  (10) Equation (13) specifies the Negbin II parameteriza-
yi ! tion of the conditional variance. i is the inverse of
When ˜ i replaces i in (10), the probability of the i and is parameterized as an exponential function,
observed count becomes conditional on the error dis- similar to the mean specification but with potentially
tribution. This conditioning can be removed, however, different variables. Since var

i  = 1/i = i , greater
by specifying the error distribution and integrating
with its probability density function to obtain the
marginal density. Most formulations specify a gamma
distribution for
variance of the error term for a given estimated mean
will result in an increase of i . I operationalize uncer-
tainty by estimating the effects of the causal vari-
i with parameter i and probabil- ables on the dispersion parameter . Variables that
ity density function g

i  as in (11) (Hausman et al. decrease  will decrease the variability and hence the
1984, Cameron and Trivedi 1986, King 1989, Long uncertainty of inventive outcomes.
1997). Although the error term can take other distribu-  
i
tions, this parameterization is flexible, computation- Var
yi x =  1 + = i + 2i  (13)
i
ally tractable, and can be derived from a variety of
assumptions. While other versions of the gamma dis- STATA estimates (12) by the method of maximum
tribution take two parameters, this derivation (from likelihood. This technique estimates both the mean i
Long 1997, p. 232) sets both to i , which forces the and variance 1/i from (12), and hence produces only
mean of
i equal to one and the variance of
i equal to a single log likelihood.

126 Management Science/Vol. 47, No. 1, January 2001

 FLEMING
Recombinant Uncertainty in Technological Search
Results
Table 4 presents estimates for the negative binomial
dispersion models of citation counts. Model 1 esti-
mates a baseline model of controls only and Model
2 adds substantive variables to the baseline model.
Piecewise models indicated a nonmonotonic effect of
component familiarity on the dispersion. To accom-
modate this, Models 3 and 4 include both first- and
second-order terms for component familiarity. The
substantive variables greatly increase the explanatory
power of Models 2 and 3, as measured by twice the
difference in log likelihoods and compared to a chi-
squared distribution with degrees of freedom equal to
the added number of variables.4 To test robustness,
Model 4 estimates substantive variables only. Coeffi-
cient magnitudes and significances vary but the sub-
stantive results remain unchanged and significant. To
further check for robustness, I split the data set in half,
randomly, and by month. Signs remained unchanged
although not always significant. The parameter esti-
mates are not standardized and should be interpreted
as the predicted multiplier effect on the mean citation
count and dispersion parameter (they should be expo-
nentiated as in (8) and (13)).
Hypothesis one proposes that usage of familiar
components will enable inventors to apply learning
from previous efforts. Such learning improves inven-
tors’ abilities to select the best components and recom-
bine them more successfully. Component familiarity
has a positive and highly significant coefficient esti-
mate. A patent receives 147% more citations at its max-
imum value of 2.28 as opposed to its lowest value
of 0.5 To interpret less extreme changes in the vari-
able, a one standard deviation increase in component
4
R-squared measures are inappropriate for maximum likeli-
hood estimates and suffer from various problems (Cameron and
Windmeijer 1996). Cameron and Windmeijer (1996) propose an
imperfect (by their argument) measure for a negative binomial
model, but it only considers the explanatory power of variables
on the mean. Cameron (in personal communication) indicated that
an acceptable R-squared measure for measuring the explanatory
power of variables on the dispersion parameter has yet to be
proposed.
5
At its maximum value of 2.28, and from the coefficient esti-
mate of 0.3958 in Model 3, the effect of component familiarity is
e
03958 ∗ 228 = 247.
Management Science/Vol. 47, No. 1, January 2001
familiarity results in an 11.3% increase in expected
citations.6
While the results support Hypothesis 1, they do
not support Hypothesis 2, that component learning
decreases the uncertainty of invention. Given that the
first-order term by itself became insignificant with-
out controls (model not shown), I estimated piecewise
models to check for a nonlinear effect. These mod-
els indicated a nonmonotonic effect, with the max-
imum negative effect at the 60th percentile of the
data. Compared to a baseline model of no effect of
component familiarity on the variance, component
familiarity decreases the variance by 9.6% at the 60th
percentile.7 Past that point, the effect becomes less
negative and eventually becomes positive at the 96th
percentile. It appears that increasing familiarity with
components has an initially negative effect on the vari-
ability of invention. Eventually this changes, however,
such that the use of very familiar components has a
positive effect on variability.
In parallel to Hypothesis 1, Hypothesis 3 argues that
reuse of familiar combinations lets inventors apply
learning from previous efforts. Such learning enables
refinement and improvement of previous inventions,
such that future inventions are more useful. Com-
bination familiarity demonstrates a positive, strong,
and highly significant effect on the mean. A patent
receives 332% more citations at combination famil-
iarity’s highest value of 11.25 as opposed to its low-
est value of 0. A one standard deviation increase of
combination familiarity results in a 10.3% increase in
expected citations.
In contrast to the contrary results of component
familiarity on the dispersion, combination familiar-
ity demonstrates predicted and highly significant
6
e
03958 ∗ 027 − 1 = 01128.
7
The maximum negative effect of component familiarity on the
dispersion (from Model 3) occurs at 0.5992. At the mean citation
count of 3.80 the variance is therefore
Var
yi x = i + 2i = 380e03958 ∗ 05992
2
+e
−04139 ∗ 05992 + 03454 ∗ 05992 
380e03958 ∗ 05992 2 = 2532
At the mean citation count of 3.80 and no effect of combination
familiarity on the dispersion at 0.5992, the variance is Var
yi x =
i + 2i = 380e03958 ∗ 05992 + e0
380e03958 ∗ 05992 2 = 2802. Hence the
maximum negative effect is
2802 − 2532/2802 = 96% less.
127
 FLEMING
Recombinant Uncertainty in Technological Search

Table 4 Negative Binomial Models of Citation Counts (All U.S. Patents, May/June of 1990)

Variable/Models Model 1 Model 2 Model 3 Model 4

Effects on the mean

Mean technology control 08830∗∗∗ 07790∗∗∗ 07800∗∗∗

00198
00215
00216
Number of prior art citations 00181∗∗∗ 00184∗∗∗ 00185∗∗∗

00012
00012
00012
Single-class dummy control −01742∗∗∗ −01568∗∗∗ −01605∗∗∗

00349
00437
00438
Number of subclasses control 00271∗∗∗ 00183∗∗∗ 00185∗∗∗

00044
00049
00049
Number of subclasses squared −00002+ −00001 −00001

00001
00001
00001
Newest subclass control −00003∗∗∗ −00009∗∗∗ −00009∗∗∗

00001
00001
00001
Number of classes control 00461∗∗∗ 00425∗∗∗ 00413∗∗∗

00097
00098
00098
Component familiarity 03807∗∗∗ 03958∗∗∗ 07195∗∗∗

00430
00441
00353
Combination familiarity 01329∗∗∗ 01301∗∗∗ 02301∗∗∗

00223
00225
00218
Cumulative combination usage −00384∗∗ −00363∗∗ −01523∗∗∗

00131
00132
00107
Constant −00899∗ −01090∗∗ −01134∗∗ 09115∗∗∗

00363
00369
00372
00230
Effects on dispersion parameter
Variance technology control −00154∗∗ −00187∗∗ −00177∗∗

00054
00060
00060
Activity control −00011 −00007 −00007

00020
00020
00020
Single-class dummy control 01001 02426∗∗∗ 02111∗∗∗

00656
00725
00731
Number of subclasses control −00170∗ −00164∗ −00178∗

00075
00081
00081
Number of subclasses squared 00003+ 00003+ 00003+

00002
00002
00002
Newest subclass control −00003+ −00003 −00002

00002
00002
00002
Number of classes control −00390∗ −00505∗∗ −00420∗

00189
00191
00193
Component familiarity 01683∗ −04139∗ −05331∗∗

00749
02048
01798
Component familiarityˆ2 03454∗∗ 03121∗∗

01135
01059
Combination familiarity −01033∗∗∗ −01011∗∗∗ −00506∗

00249
00248
00200
Constant 00545 −00177 01609∗ 01644∗

00560
00578
00819
00683
Log-likelihood −4109526 −4100192 −4099726 −4202200
∗ ∗∗ ∗∗∗
(n = 17264 all models, standard errors in parentheses, +p < 01, p < 005, p < 001, p < 0001)

128 Management Science/Vol. 47, No. 1, January 2001

 FLEMING
Recombinant Uncertainty in Technological Search
results on the dispersion. The negative effect supports
Hypothesis 4, that inventive uncertainty decreases
with the refinement of previously used combinations.
At the mean citation count of 3.80 and the mean value
of combination familiarity at 0.28, the variance is 2.2%
less than it would have been without the effect of com-
bination familiarity on the dispersion. At the maxi-
mum value of combination familiarity, the difference
grows to 64%.
Finally, Hypothesis 5 predicts that cumulative
use of a particular combination will eventually
exhaust recombinant potential. The results support
this hypothesis, as indicated by the negative and sig-
nificant estimate on cumulative combination usage.
When cumulative combination count reaches its max-
imum value of 26.76, a patent receives only 37.9% of
the citations it would have received at the variable’s
lowest value of 0. A one standard deviation increase of
cumulative combination use results in a 5.6% decrease
in expected citations.
Discussion
These results should be viewed cautiously for a
variety of reasons. The typical reservations regard-
ing the use of patent data certainly apply, most
notably that patenting practices and effectiveness vary
across industries (Levin et al. 1987). Furthermore, even
though these data cover all patented technologies
across a two-month time period, much inventive activ-
ity remains unpatented and, hence, unobserved. These
issues cause at least three problems. First, the dataset
misses unimportant inventions that failed to merit a
patent. Use of citation data mitigates this problem,
however, because it includes the bulk of relatively use-
less inventions that receive no or few citations. Sec-
ond, the dataset may miss breakthroughs that firms
chose not to patent, presumably for strategic rea-
sons. However, unless there is systematic bias in those
firms’ use of particular components or combinations,
these results should remain valid. Third, technolog-
ical communities vary in their propensity to patent.
Again, while these models controlled for much vari-
ance across technologies, they did not introduce
explicit industrial controls. But given the focus on
recombinant search across technological communities,
Management Science/Vol. 47, No. 1, January 2001
it would be difficult to compile similarly vast observa-
tions across so many communities. The dataset there-
fore trades detail and depth in exchange for breadth
and possibility of observation.
In addition to differences in the dependent variable
of citations across communities, the accuracy of the
independent variables also varies across communities.
Subclasses and combinations of subclasses represent
only a proxy for inventors’ components and architec-
tures. While the subclasses of digital hardware patents
correspond very closely to my engineering experi-
ence, other subclasses may not. For example, financial
patents tend to be classified in fewer subclasses and,
hence, may not reflect a process of recombinant search.
They may also be better characterized as knowledge or
algorithms and not technology. Such patents, however,
became popular only after the observation period.
In addition to concerns about data and variables,
these results also remain open to alternate interpreta-
tions. Most importantly, these models cannot defini-
tively separate learning and familiarity, technological
exhaustion, and life-cycle effects. For example: an
exogenous variable (such as the inventive myopia
described by Henderson (1995)) could drive the use
of particular components and combinations, such that
increased citations might simply reflect the popular
usage of those components and combinations. As a
result, the positive signs in the mean for compo-
nent and combination familiarity and negative sign
for cumulative use might reflect only the normal pro-
gression of technological life cycles, instead of learn-
ing and exhaustion. Even with the multiple controls
for differences in citation patterns across technologies,
types of combinations, and retrospective identification
of seminal subclasses, the models cannot convincingly
reject the alternative argument that the results merely
reflect the normal life-cycle progression.
These reservations do not apply, however, to the
uncertainty hypotheses. Indeed, all of the explana-
tory variables in the mean could simply be interpreted
as control variables for the dispersion estimates. The
combination result supports the intuitive argument
that uncertainty decreases with refinement and high-
lights the importance of early architectural refine-
ments as a source of destabilizing technologies for
organizations and industries (Henderson and Clark
129
 FLEMING
Recombinant Uncertainty in Technological Search
1990). In contrast, the nonmonotonic results for com-
ponents does not support the simple linear relation-
ship predicted in Hypothesis 2. It remains consistent,
however, with the argument that technological break-
throughs derive from new combinations of well-used
components (Usher 1954, Nelson and Winter 1982,
Sahal 1985). For example, Utterback (1996, p. xxvii)
argues that “radical innovations often are seen to
be based on the synthesis of well-known technical
information or components.” These results also indi-
cate that inventors take more inventive risk with
extremely familiar components. These results support
Utterback’s argument and motivate further research.
For example, are breakthroughs most likely when
inventors combine very familiar components in new
combinations? If this were the case, then break-
throughs would be most likely to emerge from social
contexts that brought together inventors with deep
experience in previously disparate fields. Such con-
texts would also be more likely to be the technologi-
cal source of potential future product life cycles and
trajectories.
Conclusion
This paper developed and tested an explanation for
the sources of purely technological uncertainty. It
argued that the source of technological novelty and
uncertainty lies within the combination of new com-
ponents and new configurations of previously com-
bined components. Inventors’ experimentation with
new components and combinations leads to less suc-
cess on average, but it also increases the variability
that can lead to breakthroughs. Empirical results sup-
ported the arguments with the exception that the use
of more familiar components has a nonmonotonic and
eventually positive effect on the uncertainty of inven-
tion. In contrast to the nonmonotonic effect of com-
ponent familiarity, the refinement of previously used
combinations has a negative and monotonic effect on
uncertainty.
Coupled with recent and complementary research
regarding market influences (Adner and Levinthal
2000) and formal economic models (Klepper 1996),
this work helps us understand the causal forces that
underlie the widely observed regularities of the prod-
uct life cycle. The data and methods presented here
130
also provide a strong basis for further investigations
into the sources of invention and technological uncer-
tainty. Most importantly, negative binomial count and
dispersion models enable researchers to analyze the
first and second moments of patent citation data. Such
models will enable us to move beyond basic counts
in analysis of patent data and quantitatively ana-
lyze the outliers of the highly skewed distributions
of inventive trials. Such tools can enable more formal
analysis of breakthrough inventions, heretofore, “the
domain of economic historians” (Scherer and Harhoff
2000). For example, the classic controversy about the
sources of technological breakthroughs, whether they
emerge from smaller, entrepreneurial, and “outside”
firms (Schumpeter 1939, Marquis 1969, Klein 1977), or
large, industrial incumbents (Schumpeter 1942), can
be reconsidered with these data and methods. The
approach presented here also has application beyond
the study of technology and patents, mainly that we
should think of varying variance as an opportunity
instead of a nuisance.
Future work should also investigate the relation-
ships between invention as a recombinant search pro-
cess and other literatures such as integrality, coupling,
and modularity (Ulrich 1995), complex systems
(Kauffman 1993), modular operators (Baldwin and
Clark 1999, Goldenberg et al. 1999), the evolution of
modular design choices (Simon 1996), and the market
implications of such issues (Christensen and Verlinden
2000). For example, interdependence between compo-
nents should increase inventive uncertainty and mod-
ularity should decrease it. Kauffman’s (1993) models
of search over interdependent landscapes also imply
a positive but nonmonotonic effect on the mean. Con-
ceptualizing technological invention as search over
interdependent landscapes implies a complexity catas-
trophe in technological evolution, as inventors face
greater numbers of components and greater interde-
pendence between them (Fleming and Sorenson 2001).
Inventors are not blind search agents, however, and
their search strategies will differ greatly and pre-
sumably be more effective than genetic recombination
(Gavetti and Levinthal 2000, Rivkin 2000).
The work has strategic implications as well. Organi-
zations that seek technological breakthroughs should
experiment with new combinations, possibly with
Management Science/Vol. 47, No. 1, January 2001
 FLEMING
Recombinant Uncertainty in Technological Search
old components. They do so, however, at the risk
of an increased number of failures. If invention is
indeed a partially blind search process, such fail-
ures are to some extent unavoidable. Simple port-
folio strategies are unfortunately not the complete
answer, because of the extreme skew of inventive
distributions. Although the overall variance of such
distributions decreases with an increasing number
of trials, firms—and indeed, as Scherer and Harhoff
(2000) demonstrate—entire economies cannot antici-
pate completely stable returns. There might exist orga-
nizational mechanisms that encourage recombinant
exploration, while limiting the downside of increased
failures. For example, firms that can screen or test
their nascent inventions more effectively will bene-
fit more and should increase the variability of their
inventive trials. The recent progress in drug design as
a result of automated screening processes provides a
salient example (Amato 1999). Finally, science should
enlighten and shorten technological search over diffi-
cult landscapes. Science can either motivate recombi-
nant search across particular technological landscapes,
or aid in search across landscapes that inventors have
discovered empirically.
Acknowledgments
The author would like to acknowledge the helpful comments of
Gautam Ahuja, Steve Barley, Bill Barnett, Phil Bromiley, Laurie
Calhoun, Mike Hannan and his seminar students, James March,
Candace Fleming, Olav Sorenson, Toby Stuart, and three very help-
ful reviewers. He would also like to acknowledge the support of
William Simpson and the Faculty Research Computing Center at
the Harvard Business School, the Department of Industrial Engi-
neering at Stanford University, and Corey Billington, Phil Johnson,
and Ellen King and the library staff of Hewlett Packard Company.
Errors and omissions remain my responsibility.
References
Abernathy, W., J. Utterback 1978. Patterns of industrial innovation.
Tech. Rev. (June) 40–47.
Adner, R., D. Levinthal. 2000. Technology evolution and
market heterogeneity: Implications for product and process
innovation. Working paper, INSEAD, France.
Ahuja, G. 2000. Collaboration networks, structural holes, and inno-
vation: A longitudinal study. Admin. Sci. Quart. 45 425–455.
Albert, M. B., F. Narin, D. Avery, P. McAllister. 1991. Direct valida-
tion of citation counts as indicators of industrially important
patents. Res. Policy 20 251–259.
Management Science/Vol. 47, No. 1, January 2001
Allen, T. 1977. Managing the Flow of Technology. MIT Press,
Cambridge, MA.
Altschuler, G. 1998. The Innovation Algorithm. Technical Innovation
Center, Inc., Worcester, MA.
Amato, I. 1999. Industrializing the search for new drugs. Fortune
(May 10)
Anderson, P., M. Tushman. 1990. Technological discontinuities and
dominant designs: A cyclical model of technological change.
Admin. Sci. Quart. 35 604–633.
Argote, L., S. Beckman, D. Epple. 1990. The persistence and trans-
fer of learning in industrial settings. Management Sci. 36(2)
140–154.
Ayres, R. 1988. Barriers and breakthroughs: An “expanding fron-
tiers” model of the technology-industry life-cycle. Technovation.
7 87–115.
Baldwin, C., K. Clark. 1999. Design Rules: The Power of Modularity.
MIT Press, Cambridge, MA.
Basalla, G. 1988. The Evolution of Technology. Cambridge University
Press, Cambridge, MA.
Bijker, W. 1987. The social construction of bakelite: Toward a theory
of invention. Bijker, Hughes, Pinch, eds. The Social Construction
of Technological Systems. MIT Press, Cambridge, MA.
Cameron, A., P. Trivedi. 1986. Econometric models based on count
data: Comparisons and applications of some estimators and
tests. J. Appl. Econometrics. 1 29–53.
, F. Windmeijer. 1996. R-squared measures for count data
regression models with applications to health-care utilization.
J. Bus. Econom. Statist. 14(2) 209–220.
Carr, F. 1995. Patents Handbook. McFarland, Jefferson, NC.
Christensen, C., M. Verlinden. 2000. Disruption, disintegration,
and the dissipation of differentiability. Working paper 00-074,
Harvard Business School, Boston, MA.
Clark, K. 1985. The interaction of design hierarchies and market
concepts in technological evolution. Res. Policy 14 235–251.
Cohen, W., D. Levinthal. 1990. Absorptive capacity: A new per-
spective on learning and innovation. Admin. Sci. Quart. 35
128–152.
Darr, E., L. Argote, D. Epple. 1995. The acquisition, transfer, and
depreciation of knowledge in service organizations: Produc-
tivity in franchises. Management Sci. 41(11) 1750–1762.
Epple, D., L. Argote, R. Devadas. 1991. Organizational learn-
ing curves: A method for investigating intra-plant transfer of
knowledge acquired through learning by doing. Organ. Sci.
2(1) 58–70.
Fleming, L., O. Sorenson. 2001. Technology as a complex adaptive
system: Evidence from patent data. Res. Policy Forthcoming.
Gavetti, G., D. Levinthal. 2000. Looking forward and backward:
Cognitive and experiential search. Admin. Sci. Quart. 45
113–137.
Gilfillan, S. 1935. Inventing the Ship. Follett Publishing Co.,
Chicago, IL.
131
 FLEMING
Recombinant Uncertainty in Technological Search
Greene, W. 1993. Econometric Analysis. Prentice Hall Inc., Upper
Saddle River, NJ.
Goldenberg, J., D. Mazursky, S. Solomon. 1999. Toward identifying
the inventive templates of new products: A channeled ideation
approach. J. Marketing Res., 36 200–210.
Hall, B., A. Jaffe, M. Trajtenberg. 2000. Market value and patent
citations: A first look. NBER Working paper 7741,
Hargadon, A., R. Sutton. 1997. Technology brokering and inno-
vation in a product design firm. Admin. Sci. Quart. 42
716–749.
Hausman, J., B. Hall, Z. Griliches. 1984. Econometric models for
count data with an application to the patents-R&D relation-
ship. Econometrica 52 909–938.
Henderson, R. 1995. Of life cycles real and imaginary: The unex-
pectedly long old age of optical lithography. Res. Policy 24
631–643.
, K. Clark. 1990. Architectural innovation: The reconfigura-
tion of existing product technologies and failure of established
firms. Admin. Sci. Quart. 35 9–30.
Jaffe, A., M. Trajtenberg, R. Henderson. 1993. Geographic localiza-
tion of knowledge spillovers as evidenced by patent citations.
Quart. J. Econom. 108 577–598.
Jorgensen, B. 1997. The Theory of Dispersion Models. Chapman and
Hall, London, U.K.
Kauffman, S. 1993. The Origins of Order. Oxford University Press,
New York.
Kim, D., B. Kogut. 1996. Technological platforms and diversifica-
tion. Organ. Sci., 7 283–301.
King, G. 1989. Event count models for international relations:
Generalizations and applications. Internat. Stud. Quart. 33
123–147.
Klein, B. 1977. Dynamic Economics. Harvard University Press,
Cambridge, MA.
Klepper, S. 1996. Entry, exit, and innovation over the product life-
cycle. Amer. Econom. Rev. 86(3) 562–583.
, 1997. Industry life cycles. Indust. Corporate Change 6(1)
145–181.
Kogut, B., U. Zander. 1992. Knowledge of the firm, combinative
capabilities, and the replication of technology. Organ. Sci. 3
383–397.
Kroeber, A. 1948. Anthropology. Harcourt Brace Jovanovich, Inc.,
New York.
Levin, R., A. Klevorick, R. Nelson, S. Winter. 1987. Appropriating
the returns from industrial research and development: Com-
ments and discussion. Brookings Paper Econom. Activity (3) 783–
831.
Levinthal, D. 1998. The slow pace of rapid technoligical change:
Gradualism and punctuation in technological change. Indust.
Corporate Change 7(2) 217–247.
Long, J. 1997. Regression Models for Categorical and Limited Dependent
Variables. Sage Publications, London, U.K.
Accepted by Karl Ulrich; received October 27, 1998. This paper was with the author 8 months for 3 revisions.
132
March, J. 1991. Exploration and exploitation in organizational
learning. Organ. Sci., 2 71–87.
, H. Simon. 1958. Organizations. Blackwell Publishers,
Cambridge, MA.
Marquis, D. 1969. The anatomy of successful innovations. Innova-
tion 1 35–48.
McCluskey, E. 1986. Logic Design Principles with Emphasis on Testable
Semi-Custom Circuits. Prentice Hall, Englewood Cliffs, NJ.
Mead, L. Carver. 1980. Introduction to VLSI Systems. Addison-
Wesley, Reading, MA.
MicroPatent. 1996. Product of micropatent USA, New Haven, CT.
Mokyr, J. 1990. The Lever of Riches: Technological Creativity and Eco-
nomic Progress. Oxford University Press, New York.
Nelson, R., S. Winter. 1982. An Evolutionary Theory of Economic
Change. Belknap Press, Cambridge, MA.
Rivkin, J. 2000. Imitation of complex strategies. Management Sci. 46
824–844.
Romer, P. 1993. Ideas and things: The concept of production is
being re-tooled. The Economist (Sept 11) 70–72.
Rosenberg, N. 1996. Uncertainty and technological change.
R. Landau, R. Taylor, G. Wright, eds. The Mosaic of Economic
Growth. Stanford University Press, Stanford, CA.
Ruttan, V. 1959. Usher and Schumpeter on invention, innovation,
and technological change. Quart. J. Econom. (November)
Sahal, D. 1985. Technological guideposts and innovation avenues.
Res. Policy 14 61–82.
Scherer, F., D. Harhoff. 2000. Technology policy for a world of
skew-distributed outcomes. Res. Policy 29 559–566.
Schumpeter, J. 1939. Business Cycles. McGraw-Hill Book Company,
Inc., New York.
, 1942. Capitalism, Socialism, and Democracy. Harper and Row
Publishers, New York.
Simon, H. 1996. The Sciences of the Artificial. MIT Press, Cambridge,
MA.
Stuart, T., J. Podolny. 1996. Local search and the evolution of tech-
nological capabilities. Strategic Management J. (Summer)
Tratjenberg, M., R. Henderson, A. Jaffe. 1997. University versus
corporate patents: A window on the basicness of invention.
Econom. Innovation New Tech. 5 19–50.
Tushman M., P. Anderson. 1986. Technological discontinuities and
organizational environments; Admin. Sci. Quart. 31 439–465.
Ulrich, K. 1995. The role of product architecture in the manufac-
turing firm. Res. Policy 24 419–440.
Usher, A.P. 1954. A History of Mechanical Invention. Cambridge, MA.
Utterback, J. 1996. Mastering the Dynamics of Innovation. Harvard
Business School Press, Boston, MA.
Vincenti, W. 1990. What Engineers Know, and How They Know It.
Johns Hopkins University Press, Baltimore, MD.
Weitzman, M. 1996. Hybridizing growth theory. Proc. Amer. Econom.
Assoc. (May) 207–212.
Management Science/Vol. 47, No. 1, January 2001